Cleaning Process

• The raw weather station data underwent a multistage cleaning and preprocessing procedure designed to ensure temporal consistency, handle missing values, and prepare the data for both linear and GNN-based models.
• The dataset is of shape (time, feature) where each node has a record for each time step resulting in each time having 8 records, one for each node.
• Data was downsampled from 1-hour intervals to 6-hour intervals to reduce noise and improve model efficiency.
• The final dataset was transformed into a 3D array off shape (time, station, feature).

Correlation Analysis

• The correlation analysis is done to minimize data leakage as many weather features are directly calculated from temperature.
• Inter-node correlation had each node was broken into its own dataset and then within the node all features were compared against each other.
• Intra-node correlationhad each feature was broken into its own dataset and then within the feature all nodes were compared against each other.

Feature Imputation

• Interpolation was incredibly important for this task as the data used in this project comes from real world weather sensors and as such is plagued with quality issues.
• For the spatial imputation step each time step was analysed and any with missing features had the missing values calculated based on the neighboring nodes and the graph edge weights.
• In situations where all nodes are missing a feature it is impossible to perform the spatial imputation step and as such a more naive temporal imputation was required.

Graph Neural Network

• The GNN is structured to model the spatiotemporal dynamics of the weather station network. It is implemented through PyTorch, the architecture is inspired by the Diffusion Convolutional Recurrent Neural Network (DCRNN) (Li et al. 2018).
• The model was tested against the last 6 months of the selected dataset.
• The GNN uses the preceding 28 time steps to forecast the next temperature value.

Linear Model

• The linear baseline model was designed as a univariate time-series regression task.
• At each time step t, the input vector aggregates the five base features across all stations:

\[ \text{tmpf}_{t+1}=\text{features}_t+\text{features}_{t-1}+...+\text{features}_{t-27} \]

Where:

\[ \text{features}_t=\text{\{tmpf, }\text{relh, }\text{sknt, }\text{drct}_{sin}\text{, drct}_{cos}\text{\}} \]

tmpf: Temperature

relh: Relative Humidity

sknt: Wind Speed

\(\text{drct}_{sin}\text{, drct}_{cos}\text{: }\)Wind Direction encoded as sine and cosine components

Data Source

Data Structure

Final Preparation

• After all preprocessing steps, the final dataset was reduced and standardized.
  • 5 features were retained: tmpf, relh, sknt, drct_sin, drct_cos.
  • 7 stations remained after filtering.
  • 4,381 time steps at 6-hour intervals (equivalent to 2 years of data) remained.

Model 1: GNN

• The Graph Neural Network was trained using the previous 28 time steps (equivalent to 7 days) and leveraged a dense spatial graph connecting all stations.
• This structure enabled the model to learn both temporal sequences and spatial diffusion patters across the weather station network.
• The model had a final MSE of 0.0562.

Model 2: LR

• The linear baseline model was trained using the same 28-time-step history with aggregated weather station data to predict the next time step’s temperature.
• All features were flattened into a single vector, treating the problem as a high-dimensional regression task with no spatial awareness.
• The model had a final MSE of 0.0147.

Key Findings

• Small Graph structures do not matter.
• Temporal Context is Crucial.
• Feature Engineering Adds Value.
• Graph Structure is Important.
• Static Graphs are Restrictive.

Key Results

• A Graph Neural Network trained on a spatiotemporal weather data may not outperform a standard linear regression baseline for short-term temperature prediction.
• Incorporating spatial structure through graph edges enabled the model to learn regional weather interactions that linear models could not.
• Careful data preprocessing, including imputation, scaling, and circular feature handling was essential to achieving strong performance from both models.
• GNNs are incredibly sensitive to parameter tuning and may outperform if provided a much larger model structure or more careful tuning.

Future Work

• These findings demonstrate the potential of traditional models compared to graph-based deep learning approaches.
  
• it is apparent that the reliance on aggregating stations as is done with the linear model most likely only worked due to the close proximity of the stations.
• we still believe there is potential in the application of a graph-based deep learning approach when it comes to large scale weather forecasting.
• If the dataset was instead made from all of the stations across the US it most likely wouldn’t be possible to aggregate in a way that still preserves spatial information.